AI Patient Education Materials: A Fact-Check and Reading-Level Checklist

Make patient handouts with AI in five steps: give the model your clinic's approved source text and nothing else, let it draft and simplify, have a named clinician check every clinical statement against the source, test the reading level, then record the version. AI writes the words; your clinicians remain responsible for the facts.

The danger is not that AI writes badly. It writes fluently, which is exactly why a wrong dose, a missing warning sign or an out-of-date recommendation slips through unnoticed. So this checklist is built around one rule: AI may rearrange and simplify what your clinic has already decided, but it never supplies the medical content itself.

Follow me on Instagram@sagnikteaches

The five gates every handout passes through

Treat each handout like a small product with a release process. Nothing reaches a waiting room, a website or a text message until it has cleared five gates, in this order. The order matters, because it is cheaper to catch a problem early than after layout and printing.

Connect on LinkedInSagnik Bhattacharya

Small clinics often assume this is too heavy. It is not: for a one-page handout the whole chain takes roughly an hour of staff time spread across two or three people, and most of that is the clinician's check, which you would want anyway.

Subscribe on YouTube@sagnikteaches

Gate 1: build a source pack before you open any AI tool

A source pack is a folder, or a single document, containing every piece of text the model is allowed to draw on. It might hold your clinic's existing leaflet, a guideline summary your clinicians already use, the manufacturer's patient information for a device, or a pathway document from your own service. Each item needs a title, an owner, and a date it was last confirmed current.

Why start here? Because an AI model asked "write a patient leaflet about managing high blood pressure" will happily produce one from whatever it absorbed in training, and you cannot trace a single sentence of it. A model asked to rewrite your text can be audited line by line.

What belongs in the pack, and what does not

• In: your clinic's own approved leaflets, clinician-written summaries, manufacturer instructions for devices you issue, and national or professional-body guidance your clinicians have named as the reference.
• Out: forum posts, blogs, search snippets, other clinics' leaflets you have no permission to reuse, and anything the AI tool itself produced earlier.
• Out: any patient-identifiable information. A handout is generic; it never needs a name or a case.

A filled-in source pack entry

Here is how one entry might look for a physiotherapy clinic writing a handout on caring for a sprained ankle in the first week. Everything is illustrative.

Notice the third row. The warning signs come from the clinic's own red-flag list, not from the model. That is the single most important thing to keep out of AI's hands.

Gate 2: a drafting prompt that cannot wander off

The prompt does three jobs: it confines the model to the pack, it sets the audience and reading target, and it tells the model to flag what it cannot support instead of filling the gap. Paste the source text below the instructions. Use a tool whose privacy terms your clinic has read (the privacy-terms checklist covers what to look for), and keep all patient data out.

You are helping a clinic turn approved source text into a patient handout.

RULES
1. Use ONLY the source text between the lines marked SOURCE. Do not add
   medical facts, doses, timings, statistics or warnings from your own knowledge.
2. If the source does not cover something a patient would expect, write
   [NOT IN SOURCE: topic] in the handout instead of guessing.
3. Audience: adults with no medical background. Short sentences, everyday
   words. Explain any medical term in brackets the first time it appears.
4. Layout: a title, one sentence on what this handout is for, headed sections
   with short bullet lists, and a box called "When to get help urgently"
   copied word for word from the red-flag text in the source.
5. After the handout, list every sentence you changed in meaning or merged,
   so a clinician can check them first.

SOURCE
[paste approved text here]
SOURCE ENDS

Rule 2 earns its keep. It converts silent invention into a visible placeholder the clinician can see at a glance. Rule 5 gives the reviewer a short list of the risky edits instead of making them reread everything with equal suspicion.

Sample output, and what you would fix

Say the model returns this fragment for the ankle handout (illustrative):

In the first few days, rest your ankle and avoid putting weight on it. Gentle movement helps recovery, so start moving your foot up and down as soon as it feels comfortable. [NOT IN SOURCE: how long to use ice] Most people recover fully within a few weeks.

Three things deserve attention. First, "rest and avoid putting weight" sits awkwardly beside "start moving your foot", so the clinician must decide which wins and when. Second, the placeholder shows the pack had no icing guidance, so someone either adds a sourced line or removes the topic. Third, "Most people recover fully within a few weeks" appears nowhere in the pack. It is a prognosis claim the model supplied itself. Delete it unless a clinician can point to a source for it.

Gate 3: the clinician's sign-off, step by step

This is the gate that makes the handout safe, so give it structure. A clinician who is simply asked "is this OK?" will skim, and fluent text invites skimming. Give them a task that forces comparison.

1. Print or open the handout beside the source pack. Two windows or two pages, not memory.
2. Highlight every sentence that makes a clinical claim: anything about what to do, how long, how much, how often, what is normal, what is dangerous.
3. Find each claim in the source. Tick it if it matches. Mark it if it was reworded, and read the reworded version for changed meaning.
4. Delete or flag anything with no source. Do not "fix it from experience" and leave it unsourced; either add the source to the pack or take the line out.
5. Read the urgent-help box character by character against the red-flag text. It must be identical in meaning and usually identical in wording.
6. Check numbers and units. Doses, durations, frequencies and measurements are where a plausible-looking slip hides.
7. Sign and date with a statement such as "Clinically reviewed against sources listed in the log".

The four errors a sign-off most often misses

These are the failure patterns worth briefing your clinicians on, because they look fine at a glance.

• Softening. "Seek urgent help" becomes "consider speaking to your doctor". The meaning of a warning has been diluted without a single wrong word.
• Merging. Two separate pieces of advice for two different groups of patients get fused into one blanket instruction.
• Over-completeness. The model adds a reassuring or tidy-sounding line to round off a section. It reads as helpful and is unsourced.
• Stale defaults. The model leans on older, more common advice where your source reflects a newer pathway. This only shows if the reviewer compares against the source, not against what feels familiar.

No diagnosis content written by the model

Set a firm boundary in your clinic policy: the model never generates content that tells a patient what they have, what to take, or what to do about a symptom. That content comes only from clinician-approved text. The model's job is limited to wording, order, layout and plain-language rewriting.

A workable policy line reads: "AI tools may be used to reformat, simplify and translate approved clinic text. They may not be used to originate clinical information, diagnoses, treatment advice, doses or triage guidance." Put it in the staff handbook and the handout log so the rule is visible at the moment of use.

The same boundary applies if a patient asks the front desk for "something on this condition". A receptionist should hand out an approved leaflet, never paste a question into a chatbot and print the reply. The front-desk safety tutorial covers where the line sits for incoming enquiries.

Gate 4: test the reading level properly

Readability scores are a screen, not a verdict. They measure sentence length and syllables per word, nothing more. A short sentence can still be unclear, and a long one can be perfectly plain. Use scores to find the paragraphs that need attention, then use a person to confirm.

Run the numbers in Word

In recent Microsoft 365 versions of Word, choose Home > Editor > Document stats. In older versions, tick Show readability statistics under File > Options > Proofing, and the statistics appear when the spelling and grammar check finishes. Word reports two scores. The Flesch Reading Ease score runs on a 100-point scale where higher is easier; Microsoft's own guidance suggests 60 to 70 for most standard documents. The Flesch-Kincaid Grade Level converts the same inputs into a school grade, and Microsoft suggests about 7.0 to 8.0 for most documents. Both rely on average words per sentence and average syllables per word.

Agree a target and write it into your policy. Health-literacy guidance usually sets the bar below Word's general default: published readability studies cite a sixth grade target from a national medical association and an eighth grade target from a national health research agency. Pick one number for your clinic, for example "Grade Level 8 or below, Reading Ease 60 or above", and apply it to every handout, with a stated exception for a few unavoidable terms such as a drug or device name.

A before-and-after rewrite

Here is a typical fix. The original sentence has 29 words and a long chain of clauses.

The second version has two short sentences, plainer words and a direct instruction. The meaning is unchanged. It is also the kind of edit an AI model does well when you ask it to "rewrite for a reader at grade 6, keep the meaning identical, change no numbers". After any rewrite the clinician checks again, because simplifying is exactly where "should" quietly turns into "can".

The lay-reader test

Ask one person who is not a clinician, such as a front-desk colleague, a volunteer or a patient-group member, to read the handout and answer three questions without help:

1. What is this handout asking me to do?
2. Which symptoms mean I should get help straight away?
3. Which word or sentence, if any, did you have to read twice?

If the reader cannot answer question two from the box alone, the box is too long or too buried. Move it to the top or the bottom of the page, set it apart, and shorten it. Record the reader's notes in the log, because they are the evidence the handout works for the people it is written for.

Other access checks that take five minutes

• Type size and contrast. One national health agency's plain-language guidance says to aim for 12 point or larger in printed materials. Decide whether your patients need bigger type than that, write the minimum into your policy, keep dark text on a plain light background, and apply it to every handout.
• Headings and lists. People scan. A handout of solid paragraphs fails the waiting-room test.
• Alternative formats. Decide whether the handout needs large print, a text-message version or a translation. A translation made by AI is a second document that needs its own check by a fluent reader, and the machine-translation tutorial explains how to do that.
• Images and diagrams. Pictures of exercises or devices come from your own materials or properly licensed sources, never from an image generator, which can invent anatomically impossible figures.

Gate 5: the version log that keeps handouts honest

A handout without a record is a liability. If a patient was harmed by outdated advice, the first question is which version they were given and who approved it. A simple spreadsheet answers that in seconds. Keep one row per version.

Two habits make the log work. First, withdraw old versions physically: clear the rack and remove the old PDF from the website the same day a new version goes live. Second, put the version and review date on the handout itself, in the footer, so anyone holding a printed copy can tell how old it is.

One knee handout through all five gates

Say a four-person physiotherapy clinic wants a handout on a common knee problem. Their clinical lead already has a two-page internal protocol, a one-page leaflet from five years ago and a red-flag list. Here is how the process runs, with illustrative times.

1. Source pack (15 minutes). The practice manager gathers the three documents and notes the dates. The leaflet is five years old, so the lead physiotherapist marks two exercise progressions as outdated and removes them from the pack. This is the step that stops the model reproducing old advice.
2. Draft (10 minutes). The manager pastes the pack into the prompt above. The model returns a one-page handout with two [NOT IN SOURCE] placeholders (return-to-sport timing and footwear) and a short list of merged sentences.
3. Clinical check (25 minutes). The physiotherapist works through the seven-step sign-off. She deletes the footwear placeholder, adds a sourced line on return-to-sport timing from her protocol, and restores a "stop and get advice" wording the model had softened to "you may wish to check with someone".
4. Readability (10 minutes). Word reports Reading Ease 52 and Grade Level 9.1, missing the clinic's target. The manager asks the model to rewrite only the sentences over 20 words, keeping meaning and numbers unchanged, then sends the result back to the physiotherapist for a two-minute recheck. Scores move to 66 and 7.0.
5. Lay reader (10 minutes). The receptionist reads it cold. She stumbles on "weight-bearing". It becomes "putting your weight on your leg".
6. Release (5 minutes). Version 1 is logged, the footer gets "v1, review due in 12 months", and the old leaflet is removed from the rack.

Total: about 75 minutes, of which 25 are the clinician's. For a handout used by dozens of patients a month, that is a fair price for something nobody has to apologise for later.

Four handout jobs that need extra care

Turning a long guideline into a one-page summary

Long source, short output is where merging errors multiply. Ask the model to extract rather than summarise: "List every instruction for patients in this document as separate bullets, quoting the sentence it came from." The clinician then selects which bullets belong in the handout, and the model only formats them. You lose a little elegance and gain a trail.

A medicines or device handout

Treat the manufacturer's patient information or the clinic's pharmacist-approved text as the sole source and instruct the model to copy dosing and storage lines verbatim. Do not let it reword numbers. Have a second person, ideally a pharmacist or the prescriber, check every figure. If your clinic cannot name that second checker, do not use AI for this category.

A text message or social post version

Shorter formats tempt people to skip the process. They should not. A text reminder about preparing for a procedure is clinical content too. Derive it from the signed-off handout, never from scratch, and log it under the same version as its parent.

An existing handout nobody can find the source for

Many clinics inherit leaflets with no author and no date. Do not feed one to AI to "tidy up". Treat it as unreviewed, ask a clinician to rebuild a source pack and approve a fresh version, and then use AI only for the wording.

Where AI-drafted handouts go wrong in practice

A realistic failure helps the team remember why the gates exist. Imagine a clinic that skipped the source pack and asked a chatbot for a handout on recovering after a minor procedure. The result looked polished, scored well on readability, and told patients they could resume normal activity within two days. A clinician skimming it saw a friendly document and signed it. In fact, the clinic's own protocol said to wait longer for patients on certain medication. Nothing in the text looked wrong. It was found only when a patient phoned with a question, and every copy already printed had to be recalled.

The lesson is not "AI is unreliable". It is that a score and a signature are worthless unless the signature means "I compared this to the source". That is why gate 3 uses highlighting and tick marks: it makes comparison the task rather than an assumption.

Who runs the process, and what it costs in time

• Owner of the process: usually the practice manager, who runs gates 1, 2, 4 and 5.
• Clinical approver: one named clinician per topic area, with a deputy. If there is no deputy, handouts stall during leave; plan for it.
• Tool cost: a general AI assistant on a paid business plan is enough; list prices change, so check the vendor's current page. Word's readability statistics are free. The version log is a spreadsheet.
• Time: allow roughly an hour per one-page handout at first, falling as the prompt, source pack habits and templates settle.

If you are weighing whether to add AI to the wider clinic, start with admin tasks where the risk is lower. The list of clinic admin tasks is a gentler place to build confidence, and a quiet trial in shadow mode lets staff practise the sign-off before any patient sees the output.

A one-page checklist to pin up

Copy this into your staff handbook, and tick it for every handout before release.

• Source pack assembled, each item dated and owned.
• No patient-identifiable information used anywhere in the drafting.
• Prompt confines the model to the pack and asks for placeholders instead of guesses.
• Every clinical claim traced to a source by a named clinician.
• Urgent-help box matches the clinic's red-flag text.
• All numbers, units and timings double-checked.
• Reading Ease and Grade Level recorded against the clinic's target.
• One lay reader answered the three questions without help.
• Translations, large print and text versions checked separately.
• Version, review date and sources logged; old version withdrawn the same day.

If your clinic is also drafting letters to other practitioners, the referral letter tutorial applies the same draft-then-check discipline to a different document. And if you would rather map your approval process with someone, that is what my AI implementation consultation covers.

Questions clinics ask before they publish an AI-drafted handout

Can AI write a patient handout from nothing?

It can, but you should not let it. A model asked for a handout on a condition fills gaps from memory, and memory is where wrong doses and out-of-date advice creep in. Give it your own approved source text and tell it to use nothing else. Then every sentence can be traced back to something your clinicians already trust.

Which reading level should a handout aim for?

Widely cited health-literacy guidance puts patient materials at about a sixth to eighth grade reading level, but your clinic should set its own target and write it down. Word reports a Flesch Reading Ease score, where Microsoft suggests 60 to 70, and a Flesch-Kincaid Grade Level, where it suggests about 7 to 8. Use both as a screen, then test with a real reader.

Do we need to tell patients a handout was drafted with AI?

There is no single answer, because rules differ by place and by regulator. What matters most is that a named clinician has reviewed and approved it, and that the handout shows who and when. Ask your professional body or compliance adviser whether any disclosure wording applies to you, and apply it consistently.

How often should handouts be reviewed?

Set a review date on every handout, for example 12 months, and bring it forward whenever a guideline, drug, device or local pathway changes. The version log makes this easy: sort it by review date each quarter and work through anything overdue. A handout with no review date is one nobody is responsible for.

Is it safe to paste patient details into ChatGPT or Claude?

Don't. A handout is generic teaching material and never needs a name, date of birth or case detail. Keep identifiable data out of the drafting tool entirely, and check the privacy terms of any tool your clinic uses before staff put anything health-related into it.

Further reads

• Is an AI Scribe Worth It for a Small Private Clinic? — Another clinic AI tool that needs clinician checking.
• AI Tools for Chiropractors: Report of Findings, Care Plans and Handouts — A sector example of care plans and handouts with AI.
• AI for Dental Hygienists: Recall Messages and Patient Education Scripts — Patient education scripts from a dental hygiene angle.
• Creating Home Exercise Programmes With AI for Physio Patients — Home exercise sheets drafted with AI and checked.
• AI Governance for a Small Business: Who Decides, Approves, Checks — Who decides, approves and checks AI output in a small team.
• AI Tools for Physiotherapy Clinics: What Each One Actually Does — What each AI tool does in a physiotherapy clinic.
• How a Clinic Can Use AI to Cut Missed Appointments — Cut no-shows with reminders that stay inside safe limits.
• AI Marketing for Medical Practices Without Making Health Claims — Service pages, review replies, referrer updates and local search for a medical practice, drafted with AI and kept free of health claims.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support page on readability statistics in Word (checked 4 November 2026); a peer-reviewed readability study citing a medical association's sixth grade and a health research agency's eighth grade targets; a national health agency's Clear and Simple guidance (type size); a public-health agency's Clear Communication Index page. Reading-level targets are guidance, not a rule.